Hover, Transition, and Level Flight Control Design for
a Single-Propeller Indoor Airplane
Adrian Frank∗, James S. McGrew†, Mario Valenti‡, Daniel Levine§, and Jonathan P. How¶
Aerospace Controls Laboratory, Massachusetts Institute of Technology, Cambridge, MA, 02139
This paper presents vehicle models and test flight results for an autonomous
fixed-wing aircraft with the capability to take off, hover, transition to and from
level-flight, and perch on a vertical landing platform. These maneuvers are all
demonstrated in the highly space constrained environment of the Real-time indoor
Autonomous Vehicle test ENvironment (RAVEN) at MIT. RAVEN promotes the
rapid prototyping of UAV planning and control technologies by allowing the use of
unmodified commercially available model aircraft for autonomous flight. Experimental results of several hover tests, transition maneuvers, and perch landings are
presented. By enabling a fixed-wing UAV to achieve these feats, we demonstrate
that the desirable speed and range performance of an autonomous fixed-wing aircraft in level flight can be complimented by hover capabilities that are typically
limited to rotary-wing vehicles. This combination has the potential to significantly
ease support and maintenance of operational autonomous vehicles.
I.
Introduction
Unmanned aerial vehicles (UAVs) are being used to perform a wide variety of missions in
an ever expanding range of operational environments. As UAV capabilities increase, so do the
possibilities for practical application in the real world. The ability to hover would give a UAV certain
flexibility with regards to operating environment and mission capabilities generally not possible
with a traditional fixed-wing aircraft design. By enabling a fixed-wing UAV to perform transitions
between hover and flight (and vice-versa) in a confined environment the desirable capabilities of
rotary-wing and fixed-wing aircraft configurations can be realized in a single design. The UAV
gains the ability to perform a multitude of missions previously limited to rotary-wing UAVs while
maintaining the high speed cruise and extended range associated with fixed-wing UAVs. In an
open environment, a fixed-wing UAV could fly a loiter pattern to remain on station for extended
time periods, however, with the added spatial constraints imposed by most urban environments,
a loiter pattern may not be feasible. Furthermore, missions such as reconnaissance, search and
rescue, surveillance, and various other scenarios can sometimes benefit from having a sensory unit
that can maintain a stationary position. Using fixed-wing UAVs for sensing missions in enclosed
spaces can provide additional operational advantages, for example, a fixed-wing UAV can perform
quick dashes from one position to another, allowing for quick relocation. The capability to hover
∗
Visiting Student, Department of Aeronautics and Astronautics, a frank@mit.edu, M.Sc. Candidate, Department
of Aeronautical and Vehicle Engineering, Royal Institute of Technology, Stockholm, Sweden
†
S.M. Candidate, Department of Aeronautics and Astronautics, jsmcgrew@mit.edu
‡
Ph.D., Department of Electrical Engineering and Computer Science, valenti@mit.edu, Student Member AIAA.
§
S.B. Candidate, Department of Aeronautics and Astronautics Engineering, denu@mit.edu.
¶
Professor of Aeronautics and Astronautics, jhow@mit.edu, Associate Fellow AIAA.
1 of 18
American Institute of Aeronautics and Astronautics
enables vertical take-offs and landings, which do not require runways or facilities with large spaces.
These combined factors make fixed-wing UAVs with transition capabilities ideal for missions where
spatial constraints are dominant.
The usefulness of transition capabilities in fixed-wing UAVs has been acknowledged by numerous
research teams. The first successful manually controlled transitions were performed in 1954 with the
Convair XFY-1 “Pogo”.14 Currently, there are several research groups working to develop vehicles
that possess the capability to fly in both a level-flight and hover configuration. A custom designed,
radio-controlled (R/C) airplane was developed at Drexel University. The airplane is manuallycontrolled in level-flight operations and transitions to a computer-controlled hover configuration
upon user input. The airplane has onboard processing and sensing using a commercially available
internal measurement unit (IMU).13 At Brigham Young University (BYU) a trajectory generator
has been used to simulate the actions of a commercially available R/C model of the Convair XFY1 to study the autonomous hover, flight, and transitions problem.8 Researchers at University of
Sydney in Australia have designed and built a T-wing tail-sitter currently capable of autonomous
hover.9 An on-line video presents the vehicle’s capability to autonomously take-off from a vertical
orientation, hover, transition to conventional flight, fly a path according to predefined waypoints,
transition back to hover, and land. The vehicle uses onboard processing and sensing including GPS
measurements for position sensing.11 Successful autonomous transitions from flight to hover (and
vice-versa) have also been performed by researchers at Georgia Tech. Their experiments used a
large, commercially available, R/C airplane equipped with appropriate electronics.3
Although transitions from a traditional aircraft level-flight mode to the hover position (and
vise-versa) have been discussed as a means of enabling missions to be performed in enclosed areas,
the projects discussed above have all been undertaken in areas where space has not been a limiting
factor. In previous research, autonomous take-offs and landings are generally performed either
from a flat, open surface, or by hand-launch. This effectively limits the possibility to launch and
land in an environment where space is limited or access is difficult. In addition, to the best of our
knowledge the first reported instance of a fixed-wing autonomous UAV perching on vertical landing
platform is shown in [12]. However, flight test data for this landing experiment was not provided
in this video submission.
This paper presents both vehicle models and test flight results for an autonomous fixed-wing
airplane that is designed to take-off to a hover, transition to and from traditional fixed-wing level
flight modes, and perch on a vertical landing platform in a space-constrained environment. These
test flights were performed indoors, thus introducing critical spatial constraints comparable to
elements of an urban environment. A model of the fixed-wing airplane in both hover and level flight
is presented, followed by descriptions of the control systems and guidance logic. In addition, this
paper presents flight and perching results for a fixed-wing aircraft in hover configuration. Flight test
results for the vehicle flying in the transition configuration and level flight mode are also provided.
All of the flights provided in this paper were performed using the MIT Real-time indoor Autonomous
Vehicle test ENvironment (RAVEN) and are available online at http://aerobatics.mit.edu
II.
Hardware and Control Structure
The first objective of the project was to find a vehicle platform that could be used to study and
implement technology that allows a fixed-wing airplane to hover, land/perch, transition and fly in
an indoor environment and other heavily constrained spaces. Two different commercially-available
off-the-shelf (COTS) R/C fixed-wing aircraft were used in these tests. First, the Blade 3D foam
airplane was used for early flight testing due to its lightweight, low-cost and, durable airframe.
Later, the Ikarus Yak 54 Shock Flyer foam airplane was selected for the final flight testing due to
its better aerodynamic behavior and higher structural rigidity. Both airplanes have the capability
2 of 18
American Institute of Aeronautics and Astronautics
to make the tight turns necessary to enable level flight and transitions in RAVEN’s constrained
flight space.
Using the architecture described in [7], each air vehicle can be controlled by a ground-based computer via the trainer port on an R/C transmitter. In addition, the attitude and position of the vehicle can be sensed
by a motion capture system by placing lightweight reflective markers on
the vehicle. The markers reflect visible red light emitted from LED-arrays
mounted on the cameras (see Figure 1). Images from several cameras are
compared and the position and attitude of the vehicle is calculated based
on a known geometry of markers on the vehicle. Therefore, all computations are done offboard, allowing the use of the vehicle in its off-the-shelf
configuration, which enables rapid prototyping and cost effective testing. Figure 1. VICON motion
If desired, this architecture could be augmented by an onboard internal capture camera. Lens
at center of the
measurement unit (IMU) and processor for further indoor or outdoor au- located
LED-array.
tonomous operations.
As shown in Figure 2, the control structure used for the airplane is that of an automaton with
three states: land/perch, hover, and flight. Similar structures have been used in the past to control
rotary-wing vehicles performing complex maneuvers with success.4 Hover is defined as a state in
which the airplane maintains a steady or semi-steady position with the aircraft’s nose perpendicular
to the ground. The airplane can be given movement commands in hover. Level flight is a state
in which the airplane’s wings generate the main portion of the lift and the nose of the airplane is
parallel to the ground.
Figure 2. Automaton structure of the control system logic.
In each of our experiments, the airplane was able to take off horizontally from the ground or
vertically from a perching stand before entering the hover state. While in hover, the airplane can
be moved to any point in space and rolled to a specified orientation about the aircraft longitudinal
axis. In addition, the vehicle can transition to and from the level flight mode, allowing the vehicle to
move quickly between desired locations. The level flight state is maintained for as long as waypoints
are available. As the airplane approaches the last defined level-flight waypoint, a transition to hover
is initiated, causing the airplane to stay over the last waypoint until commanded otherwise. These
three main parts of the control system are discussed in detail in this paper.
III.
Quaternion Attitude Description
Modeling: In this paper, it is assumed that the Earth reference frame, (xE , yE , zE ), is inertial.
In addition, there exists a set of conversions between the Earth frame and the body frame, (xB ,
yB , zB ), affixed to the airplane, assumed to be rigid. An illustration of the reference system used
can be found in Figure 3.
One of the true challenges of performing flights involving transitions is to have a global attitude
reference that allows the vehicle’s attitude to be described in a non-singular manner regardless of
3 of 18
American Institute of Aeronautics and Astronautics
the attitude of the vehicle. There are several methods of describing the attitude of an airplane, the
most common being the Euler angle representation. A primary benefit of using Euler angles is that
the angles intuitively correspond to roll, pitch, and yaw. However, a deficiency inherent in the Euler
angle description becomes clear as a vehicle approaches a pitch angle of ±90◦ , when the Euler angle
description fails to distinguish between roll and yaw.1 This singularity does not pose a problem as
long as the vehicle’s flight envelope is limited in pitch; as is the case in typical rotary-wing aerial
vehicles where large pitch angles are undesirable during nominal operation. The assumption of
a limited range of pitch angles also applies to most airplanes, where relatively straight and level
flight is the only condition of interest. For this project, both hover and flight modes of an airplane
are of interest. Therefore, a pitch angle-limited flight envelope is unfeasible, and a singularity-free
description of an arbitrary attitude is necessary.
To avoid the singularity exhibited by an Euler
angle attitude description, quaternions are used as
a global attitude reference.6 A quaternion is a 4tuple of real numbers of unit length, which limits
it to three degrees of freedom.6 A quaternion can
also be written as a rotation by an angle ζ/2 about
a three-dimensional unit-vector Ē, such that


q
0
"
#


cos (ζ/2)
 q 
q=
= 1 
 q2 
Ē sin (ζ/2)
Figure 3. Reference frame and body frame
q3
In addition, we define the zero-quaternion q (0) = [1, 0, 0, 0]T to be a reference attitude corresponding
to an Euler angle description of φ = θ = ψ = 0◦ . Where φ, θ, and ψ are rotations about xB , yB ,
and zB applied in reverse order, that is, starting by rotating about zB with the angle ψ. The
zero-quaternion here refers to the attitude in which the nose of the airplane is pointed straight
North, with the airplane in level flight.
Two key concepts are used in the following: quaternion multiplication and quaternion splitting.
A quaternion can be thought of as a rotation from the zero-quaternion to the current attitude, which
can be represented as a matrix rotation about a vector. Quaternion multiplications are successive
rotations applied to a zero-quaternion that are non-commutative, and possess laws similar to those
of matrix multiplications. Quaternion multiplication is defined as5, 6
 
1
 
 0 
p ? q = M (p) · q = M (p) · M (q) ·  
(1)
 0 
0
where p and q are quaternions and M (p) and M (q) are quaternion matrices, defined as



M (η) = 


η0 −η1 −η2 −η3

η1 η0 −η3 η2 

η2 η3
η0 −η1 
η3 −η2 η1
η0
The inverse of the quaternion matrix also follows the rules of matrix multiplication. Since quaternions have a unit length, no scaling occurs and M −1 = M T .
4 of 18
American Institute of Aeronautics and Astronautics
In controlling a vehicle, the current attitude q current is not necessarily of interest, but rather
the deviation, or error, q dev from a desired reference attitude q ref . A deviation quaternion can be
defined using quaternion multiplication as
q current = q dev ? q ref .
(2)
Given this definition of a deviation from the reference state, a quaternion split is specified as
q dev = q x ? q y ? q z .
(3)
The three quaternions resulting from (3) are defined as separate rotations about the xB -, yB - and
zB -axis, respectively, applied in reverse order. Consequently, these quaternions represent the rotations performed in an Euler angle description, where components of quaternions q x = [q0x , q1x , 0, 0]T ,
q y = [q0y , 0, q2y , 0]T , and q z = [q0z , 0, 0, q3z ]T are defined as
(1 − 2 q12 + q22 )/q0y ,
q
2
1 − 4 q0 q2 − q1 q3 ,
=
q0x = cos(φ) =
q1x = sin(φ) = (2 q0 q1 + q2 q3 )/q0y ,
q0y =
q2y =
q0z
cos(θ)
= cos(ψ) =
q3z
(1 − 2 q22 + q32 )/q0y ,
sin(θ)
=
2 q0 q2 − q1 q3 ,
= sin(ψ) = (2 q1 q2 + q0 q3 )/q0y .
Although split quaternions present an intuitive way of working with quaternions, it should be noted
that the split defined by (3) is not singularity free. In particular, since the split quaternions in
(3) correspond to Euler angles, the singularity at a pitch of ±90◦ is re-introduced. However, in
the control laws used in this paper, the quaternion split is only used for control error, or more
specifically, the quaternion deviation as defined by (2).
Throughout the flight envelope, a global quaternion based reference frame is used to keep track
of the attitude of the aircraft. The attitude of the aircraft using this global reference is denoted
q current in (2), which can be regarded as the deviation from the zero-quaternion. During hover, the
quaternion deviation is computed using the current desired reference attitude (essentially vertical).
During the transition from hover to level flight, the desired reference attitude is still given as a
vertical attitude, but a position offset is commanded to the waypoint follower to force the vehicle
to pitch the nose down (see Section VI for a detailed discussion). To ensure that the magnitude
of the pitch error will always be less than 90◦ during a transition from hover to level flight, the
controller mode changes into the level flight state as the pitch angle drops below 30◦ (pitch angle
of 0◦ corresponds to an attitude where the longitudinal axis of the airplane is in the horizontal
plane). The desired reference of the level flight controller is then based on the requirements to
maintain horizontal flight (again see Section VI for details). Note that in the event that the pitch
error should exceed 60◦ , the controller will change from the level flight state to hover. Due to the
potential hazard of an arbitrary transition to level flight, the airplane always remains in the hover
state until a transition is enabled by the operator.
IV.
Airplane in Hover
Modeling: Since the airplane used in this research has only one propeller, a throttle input will
produce both thrust and a moment due to rotation of the propeller. The propeller moment comes
from two main contributors. The first contributor is the moment created by a change in propeller
speed. This propeller acceleration moment is calculated as
Macc = Iprop ω̇,
5 of 18
American Institute of Aeronautics and Astronautics
(4)
where Iprop is the moment of inertia of the propeller defined in the xB -direction. The moment Macc
about the airframe xB -axis is created by the angular acceleration of the propeller, ω̇.
Propeller drag is the second contributor to the total propeller moment and can be estimated
using the NACA-standardized thrust and power coefficients CT and CP .2
CT
=
CP
=
T
ρn2 d4
P
ρn3 d5
(5)
(6)
In equations (5) and (6) the thrust T = mg is determined using Newton’s second law, where
the thrust is the force necessary to keep the airplane in a sustained hover, that is, the mass of
the airplane m times the gravitational acceleration g. In this derivation, it is assumed the thrust
generated by the propeller is the only force counteracting gravity. The flow over the propeller is
assumed to be essentially incompressible, and hence, the density of the air ρ is constant. The
remaining parameters are the diameter of the propeller d, and the rotational speed of the propeller
in revolutions per second, n. With a known thrust coefficient, the rotational speed of the propeller
can be estimated using (5). The moment created by propeller drag can then be estimated using
the relationship for power, P = Mdrag 2πn. The moment can be expressed as
Mdrag =
T dCP
.
2πCT
(7)
During hover, the control surfaces are only partially covered by the airflow from the propeller.
Utilizing actuator disk theory, the velocity of the propeller airflow at the control surfaces can be
estimated. With the additional assumptions of steady flow, discontinuous pressure, and continuous
velocity, the velocity of the propeller airflow at the control surfaces in hover can be estimated using
Bernoulli’s equation as
s
2T
uprop =
,
(8)
ρAdisk
where uprop is the axial velocity of the propeller airflow, and Adisk = πd2 /4 is the area of the
actuator disk representing the spinning propeller.
The lift coefficient derivatives for the ailerons, elevator, and rudder, CLδail , CLδele , and CLδrud ,
are determined using flat plate approximations of the control surfaces. These approximations are
predicated on airplane geometry. Lift produced by the wings can be calculated as
1
Lwing = ρu2∞ Awing CLα α.
2
(9)
Again, a flat plate approximation of the wing lift coefficient derivative CLα is valid because of airfoil
geometry. The angle of attack can, using small angle approximation, be described as α = żB /u∞ .
While in a hover, aileron deflection does not greatly affect roll. Although the deflection of the
ailerons necessary to maintain a desired heading is considerable, a significant amount of drag is
produced. This drag contribution can be modeled as
Dail
2 CL2 δ δail
1 2
ail
= CD,0 +
ρu
Aaileron,prop .
πe
2 prop
A
(10)
As a pitching or yawing motion is induced to the airframe, the propeller airflow creates an
apparent angle of attack of the control surfaces. This apparent angle of attack in turn causes
6 of 18
American Institute of Aeronautics and Astronautics
damping moments in the aircraft pitch and yaw dynamics. The damping due to propeller airflow
lag can be modeled as
qdamp = Lelevator αelevator Aelevator
rdamp = Lrudder αrudder Arudder
αelevator =
αrudder =
Celed
=
Crudd
=
1 2
ρu
2 prop
1 2
ρu
2 prop
Lprop2elevator q
uprop
Lprop2rudder r
uprop
qdamp
quprop
rdamp
ruprop
(11)
(12)
(13)
(14)
(15)
(16)
where qdamp and rdamp are the pitch and roll damping moments, Lprop2elevator and Lprop2rudder are
the distances from the propeller to the elevator and rudder, and αelevator and αrudder are the apparent
angles of attack, respectively. Separating variables from constants, the dimensional elevator and
rudder drag constants Celed and Crudd can be defined as above.
Based on the model developed so far, the general equations of motion for an airplane can be
expressed using a quaternion attitude description as
h
i
xE = 2 q1 q2 + q0 q3 xB + 1 − 2 q12 + q32 yB + 2 q2 q3 − q0 q1 zB
h
i
yE = 1 − 2 q22 + q32 xB + 2 q1 q2 − q0 q3 yB + 2 q0 q2 + q1 q3 zB
(17)
i
h
zE = −2 q1 q3 − q0 q2 xB − 2 q0 q1 + q2 q3 yB − 1 − 2 q12 + q22 zB
C2 1 2
1
Dail
ẍB = 2 q1 q3 − q0 q2 g + δthrottle − CD,0 + L
ρu∞ Awing −
m
πe
2
m
1 2
2
ÿB = 2 q0 q1 + q2 q3 g + CLδrud ρ uprop Arudder,prop + u∞ Arudder δrudder
(18)
2
h
i
1 z̈B = 1 − 2 q12 + q22 g + CLδele ρ u2prop Aelevator,prop + u2∞ Aelevator δelevator − Lwing
2
A
1
Iy − Iz
1 ṗ = qr
+ CLδail ρ u2prop Aaileron,prop Laileron,prop + u2∞ Aaileron Laileron
δaileron
Ix
2
Ix
Macc + Mdrag
(19)
−
Ix
L
Celed uprop q
Iz − Ix
1 elevator
q̇ = pr
+ CLδele ρ u2prop Aelevator,prop + u2∞ Aelevator
δelevator −
Iy
2
Iy
Iy
Ix − Iy
Crudd uprop r
1
Lrudder
ṙ = pq
+ CLδrud ρ u2prop Arudder,prop + u2∞ Arudder
δrudder −
Iz
2
Iz
Iz
7 of 18
American Institute of Aeronautics and Astronautics
q̇0 =
q̇1 =
q̇2 =
q̇3 =
1
− pq1 − qq2 − rq3
2
1
pq0 − qq3 + rq2
2
1
pq3 + qq0 − rq1
2
1
− pq2 + qq1 + rq0
2
(20)
where the vector [q0 , q1 , q2 , q3 ]T is the attitude quaternion. The variables Aaileron,prop , Aelevator,prop ,
Arudder,prop , Aaileron , Aelevator , and Arudder are the areas of the control surfaces covered by the propeller airflow and the total areas of the control surfaces affected by freestream airflow, respectively.
The freestream airflow velocity is conventionally denoted as u∞ . The lengths Laileron,prop , Laileron ,
Lelevator , and Lrudder are the moment arms of the various control surfaces.
C2
Drag is calculated using the induced drag coefficient CDi = πeL and parasitic drag coefficient
CD,0 , where CL = CLα α is the lift coefficient of the wing, e is the wing efficiency, and
is the
aspect ratio of the wing. The drag acting on the airplane while in a hover condition is negligible.
This assumption can be made since the freestream velocity is very low. The force terms in the yB and zB -axis due to control surface deflection are small in hover, again due to very low freestream
velocity, and can hence be disregarded.
Due to symmetry of the airframe, the cross-coupled inertia terms Ixy = Iyz = 0. Additionally,
the term Ixz is considerably smaller than diagonal terms of the inertia tensor (Ix , Iy , and Iz ) and is
hence disregarded. Since the moment of inertia of the propeller Iprop is considerably smaller than
Ix , the propeller torque due to change in rotational speed of the propeller is negligible even when
rapid changes in motor speed occur. All rotational speeds p, q, and r are expected to be small
enough such that products thereof can be disregarded without impacting the model accuracy.
The speed of the axial propeller airflow is considered constant while the airplane is in a hover
state. It is also known that u∞ = 0 when the vehicle is stationary, and therefore
A
δaileron =
A
2
Mdrag + δ̂aileron
CLδail ρu2prop Aaileron,prop Laileron,prop
δthrottle = mg + Dail + δ̂throttle
(21)
(22)
where δ̂throttle is a small deviation from the expected throttle input necessary to maintain hover.
The contribution of aileron drag to throttle setting will always be considerably less than that of
gravity. Consequently, the aileron drag can be neglected in the horizontal movement terms.
A deviation quaternion can be calculated by projecting the current quaternion onto the product
of a heading quaternion q hdg = [cos(hdg/2),√sin(hdg/2),
0, 0]T = [c̄, s̄, 0, 0], where hdg is the heading
√
2
2
reference, and a vertical quaternion q v = [ 2 , 0, 2 , 0]T . The deviation quaternion represents the
attitude error from the reference quaternion (in this case the vertical quaternion) to the attitude
in which the airplane is currently hovering.
q current = q dev ? q hdg ? q v

c̄ q0dev −
√ 
2  s̄ q0dev +
=

2  c̄ q0dev −
− s̄ q0dev −
(23)
s̄ q1dev −
c̄ q1dev +
s̄ q1dev +
c̄ q1dev +
c̄ q2dev +
s̄ q2dev +
c̄ q2dev −
s̄ q2dev +
s̄ q3dev
c̄ q3dev
s̄ q3dev
c̄ q3dev





(24)
The deviation from the desired state is expected to be small. Consequently, all products of deviated
quaternion components other than q0dev can be disregarded (as q0dev is near unity when deviation is
8 of 18
American Institute of Aeronautics and Astronautics
small). Using deviated quaternions, the simplified quaternion model can be written as
1
δ̂throttle
m
= gq3dev
ẍB =
ÿB
z̈B =
(25)
−gq2dev
1
1
ṗ = CLδail ρu2prop Aaileron,prop Laileron,prop δ̂aileron
2
Ix
Celed uprop q
1
Lelevator
q̇ = CLδele ρu2prop Aelevator,prop
δelevator −
2
Iy
Iy
Crudd uprop r
1
Lrudder
ṙ = CLδrud ρu2prop Arudder,prop
δrudder −
2
Iz
Iz
iT
h
q̇ dev = 0 p2 2q 2r
(26)
(27)
The simplified model can be separated into two fourth-order and two second-order decoupled linear
time-invariant (LTI) systems. The fourth-order systems describe the dynamic behavior of pitch and
zB -position, and yaw and yB -position, respectively. The pitch-zB loop is affected by the elevator
input, whereas the yaw-yB loop is affected by rudder input. The second order systems describe the
xB loop and roll loop, respectively. Here, throttle controls altitude and the ailerons control roll. It
can be noted that when the airplane is near hover, and has a heading reference of 0◦ , the system
maps approximately to xE = yB , yE = zB , and zE = xB .
Hover Control: The hover controller uses full state feedback from the sensing system, which
is made possible by accurate measurement data being available at 100 Hz.7 In order to improve
position-keeping, integrators are incorporated into all four control schemes. The controller gains
are optimized using linear quadratic (LQ) control techniques. Large penalties are assigned to pitch,
roll, and their respective derivatives to ensure that the vehicle maintains a hover attitude. Small
penalties on the position deviations prevent the vehicle from making sudden movements should the
position change. The penalty on altitude deviation is large to ensure that the airplane accurately
maintains an altitude, facilitating perching.
Due to the low structural rigidity of foam airplanes, quick control surface actuation will cause
the airframe to twist. This twisting moment can be perceived by the motion capture system as a
movement of the airplane, adding noise to the position and attitude measurements. This undesirable
effect is partially mitigated by the fact that the servos cannot instantly deflect the control surfaces.
Since the servo time constants are small enough to not completely resolve the issue, the LQ design
is focused on minimizing rapid changes in control surface position. The use of the Ikarus Yak 54
Shock Flyer also helped to reduce the twisting effects.
V.
Level Flight
Modeling: The level flight controller could use the global quaternion attitude reference for determining the aircraft orientation in a manner similar to that of the hover controller. However, it
is simpler and more intuitive to develop the level flight model and controller using the standard
Euler angle representation. The Euler angles are calculated from the global quaternion attitude
reference to ensure consistency between the attitude descriptions used for hover and level flight.
The conversion is straight forward and can be found in [6]. In order to control the vehicle, we first
9 of 18
American Institute of Aeronautics and Astronautics
introduce the full equations of motion of the airplane in Euler angles.
xE = cos θ sin ψ xB + cos φ cos ψ + sin φ sin θ sin ψ yB
+ − sin φ cos ψ + cos φ sin θ sin ψ zB
yE =
cos θ cos ψ xB + − cos φ sin ψ + sin φ sin θ cos ψ yB
+ sin φ sin ψ + cos φ sin θ cos ψ zB
(28)
zE = sin θ xB − sin φ cos θ yB − cos φ cos θ zB
C2 1 2
Dail
1
δthrottle − CD,0 + L
ρu∞ Awing −
m
πe
2
m
1 2
2
= sin φ cos θ g + CLδrud ρ uprop Arudder,prop + u∞ Arudder δrudder
2
1 2
= cos φ cos θ g + CLδele ρ uprop Aelevator,prop + u2∞ Aelevator δelevator − Lwing
2
A
ẍB = − sin θ g +
ÿB
z̈B
(29)
1
Iy − Iz
1 + CLδail ρ u2prop Aaileron,prop Laileron,prop + u2∞ Aaileron Laileron
δaileron
ṗ = qr
Ix
2
Ix
Macc + Mdrag
−
(30)
Ix
L
Celed uprop q
1 Iz − Ix
elevator
+ CLδele ρ u2prop Aelevator,prop + u2∞ Aelevator
δelevator −
q̇ = pr
Iy
2
Iy
Iy
L
Ix − Iy
C
u
1 2
rudd prop r
rudder
ṙ = pq
+ CLδrud ρ uprop Arudder,prop + u2∞ Arudder
δrudder −
Iz
2
Iz
Iz
φ̇ = p + q sin φ + r cos φ tan θ
θ̇ = q cos φ − r sin φ
ψ̇ =
(31)
1
q sin φ + r cos φ
cos θ
Unlike an airplane in hover, the airplane in level flight mode follows a predefined trajectory.
Hence, a description of position derivatives (ẍB and ÿB ) does not necessarily clarify the airplane’s
dynamic behavior in the level flight condition. Therefore, a modification of the hover model of
the airplane (17 - 20) is required for level flight control. The trajectory follower used during level
flight provides the controller with a lateral acceleration necessary to keep the vehicle on the desired
trajectory. Due to the provided reference, a more intuitive way of considering the airplane dynamics
is by looking at forward velocity vxB , altitude zE , and lateral acceleration alat rather than a position
in the earth frame of reference.
For level flight, the pitch angle can be approximated using small angle approximations for
sufficiently large velocities, i.e. not a stalled configuration. The pitch angle necessary to maintain
an altitude in level flight, that is, to create sufficient lift to counteract the gravitational force on
the airplane, is
θflight =
2mg
ρu2∞ CLwing Awing
+ δthrottle
.
(32)
Based on airplane geometry, the wing lift coefficient CLwing can be estimated using flat plate approximation. It can furthermore be assumed that the aileron deflection will be small during level flight,
due to the increased airflow over the ailerons and the increase in wetted aileron area. The reduction
10 of 18
American Institute of Aeronautics and Astronautics
in aileron deflection compared to hover aileron deflection renders the aileron drag component Dail
unnecessary.
Now we define
Macc + Mdrag − qr Iy − Iz
+ δ̃aileron
δaileron = 2
CLδail ρ u2prop Aaileron,prop Laileron,prop + u2∞ Aaileron Laileron
−pr Iz − Ix
δelevator = 2
+ δ̃elevator
CLδele ρ u2prop Aelevator,prop + u2∞ Aelevator Lelevator
−pq Ix − Iy
+ δ̃rudder
δrudder = 2
CLδrud ρ u2prop Arudder,prop + u2∞ Arudder Lrudder
C2 1 2
δthrottle = m CD,0 + L
ρu Awing + 2mg sin θ + δ̃throttle
πe
2 ∞
A
Altitude changes of the airplane are made mainly by varying the pitch angle of the airplane
and thereby pointing the velocity vector slightly upward or downward. Due to this behavior, the
altitude is a function of forward velocity and pitch angle as
żE = vxB sin θ − θflight .
The forward velocity during level flight can be considered a semi-steady state, where different
velocities are possible, but with few changes of reference velocity. Hence, we can define vxB = u∞ .
Note that this relationship is used only for altitude control. The assumption of a semi-steady
velocity also promotes the description of uprop as a set speed, as well as a small angle approximation
for θ.
In the same manner as the forward velocity, the roll angle can be considered a semi-steady state.
Hence, a desired roll angle φref and desired pitch angle θref can be defined and used to determine
the equations of motion for set roll and pitch angles.
The simplified equations of motion for the airplane in level flight can be written as
v̇xB
=
alat =
żE =
ṗ =
q̇ =
ṙ =
φ̇ =
1
δ̃throttle
m
q sin φref + r cos φref u∞
θ − θflight u∞
1
1 CLδail ρ u2prop Aaileron,prop Laileron,prop + u2∞ Aaileron Laileron
δ̃aileron
2
Ix
L
Celed uprop q
1 elevator
δ̃elevator −
CLδele ρ u2prop Aelevator,prop + u2∞ Aelevator
2
Iy
Iy
Crudd uprop r
1
Lrudder
CLδrud ρ u2prop Arudder,prop + u2∞ Arudder
δ̃rudder −
2
Iz
Iz
p + q sin φref + r cos φref θref
(33)
(34)
(35)
θ̇ = q cos φref − r sin φref
The equation for the heading or yaw ψ is ignored, since the heading compensator is designed into
the guidance logic.
Flight Control: The model of the airplane in flight can again be divided into four LTI systems.
However, the systems are not fully decoupled. Depending on the roll angle, the rudder and elevator
will influence pitch and lateral acceleration to different extents respectively. Forward velocity of
the aircraft is influenced solely by the throttle. Similarly, roll is controlled by the ailerons.
11 of 18
American Institute of Aeronautics and Astronautics
Full state feedback with an integrator is used to control the forward velocity of the airplane vxB
using throttle. The zE -control loop uses full state feedback with an integrator term to produce a
desired pitch angle θref . Depending on the roll angle of the airplane at a specific instant, a mixing
of elevator and rudder is used to get the airplane to the desired pitch attitude. Lateral acceleration
is commanded by the trajectory follower, which is detailed in Section VII. The lateral acceleration
is produced by a mixing of elevator and rudder in the same manner as the pitch angle. The two
contributions to elevator and rudder deflection are then aggregated to produce the true commands.
The lateral acceleration reference acmd is also used to produce a desired roll angle as
acmd
φref = arctan
g
With the assumption that the airplane is not ascending or descending rapidly, this reference will
put the airplane in an attitude where only the elevator will be used to achieve the desired pitch
angle and lateral acceleration, that is, in an attitude corresponding to what would be considered
normal flight, where the acceleration along the yB -axis is zero. All loops are controlled using a PID
controller structure, with the exception of the lateral acceleration loop. This loop is controlled with
a proportional gain only, since the trajectory follower itself emulates a PD controller.
VI.
Transition Logic
Transitioning the aircraft between flight regimes allows the vehicle to take advantage of its
level flight capabilities and benefits (range, speed, etc.) while enabling the vehicle to operate and
land in relatively small spaces using its hovering capabilities. However, enacting this transition
can be difficult in a small space. For example, in order to transition from a hover to level flight,
the controller must adjust the vehicle’s throttle so as to allow the gravitational force acting on
the airplane to pull the vehicle’s nose down in a controlled descent, while trimming out its control
surfaces to ensure that the vehicle has the heading, pitch, and roll angles to complete the transition
successfully. Although executing transitions between flight modes is possible for a well-trained
human pilot in large open spaces, transitions must be performed very efficiently in small spaces to
enable the vehicle to recover both altitude and position control. This section describes how the
transitions are performed automatically.
Transition to Flight: In order to transition to level flight in a confined space, the airplane
needs to quickly gain speed while losing a minimal amount of altitude. This is achieved by using the
hover controller. With an instantaneous movement of the reference waypoint to a distant location,
in the direction of the desired transition, these goals can be met satisfactorily. As the airplane tries
to move quickly to the new location, the pitch angle is reduced. By selecting the distant waypoint
15 m away from the hover location, the loss of altitude is small, while the airplane gains enough
speed to meet the requirements of horizontal flight. The distant waypoint is used only to produce
a rapid decrease in pitch angle; the controller switches into the level flight state after traveling only
about 1 m, where the reduction of pitch angle to less than 30◦ required to switch is achieved. It
was also found during flight tests that a 5% reduction in throttle from that required to maintain
altitude during hover helped the airplane pitch forward slightly quicker. This improvement can be
attributed to the propeller airflow lag that causes the airplane to resist the desired forward pitching
moment. A larger reduction in throttle causes an even quicker drop in pitch. However, if the pitch
of the airplane is brought down faster, the airplane loses more altitude. Hence, a balance between
altitude loss and pitch rate has to be achieved.
Transition to Hover: The challenge of transitioning to hover is obviously the opposite of the
challenges of transitioning to flight. As the airplane transitions to hover, altitude needs to be
12 of 18
American Institute of Aeronautics and Astronautics
maintained, and the forward velocity needs to quickly go from level flight velocity to zero. In order
to achieve the desired transition characteristics, the elevator and rudder are used in combination
to minimize the time used to pitch the airplane to a nose up attitude. This rapid procedure puts
the vehicle in a high angle of attack situation throughout the maneuver. Due to the dynamics of
the airplane, a major portion of the forward velocity is lost to drag. To prevent the airplane from
entering a complete stall, which would cause the aircraft’s nose to pitch forward and compromise
the desired transition, the throttle is set slightly lower than what is necessary to keep the airplane
in a hover. This throttle setting ensures that full control authority is maintained for elevator and
rudder by providing propeller airflow over the control surfaces. At the same time, most of the wing
lacks coverage by the propeller airflow, which causes high drag from the stalled portions of the wing
to quickly reduce the speed of the airplane. During transition to hover, the ailerons work in the
same manner as during regular flight, that is, the ailerons are used to keep the airplane along the
predefined trajectory leading up to the hover reference point.
The transition logic used to transition the airplane from level flight to hover is characterized by
δelevator = cos(φ) δelevatormax
CLδrud Arudder Lrudder
Iz
CLδele Aelevator Lelevator
Iy
δrudder = sin(φ) δruddermax
(36)
(37)
δthrottle = k mg,
(38)
where k is a number slightly lower than unity. The aileron drag component is not included here,
since the contribution to the total throttle is negligible in comparison with mg. Altitude hold is
also not of significant interest during transition from level flight to hover.
As can be seen, transition to hover is done in an open-loop manner for all controls except for
ailerons. The scaling on elevator is based on the assumption that the airplane will have a larger
elevator control authority than rudder control authority. This is based on the fact that the elevator
control surface is larger than the rudder control surface, while the maximum deflections of each
is similar. This assumption was true for both airplane models used in our flight testing. The
scaling factor ensures that the combination of elevator and rudder creates a pure pitching moment
regardless of the roll angle of the airplane during transition to hover.
VII.
Vehicle Guidance Logic
The trajectory following logic is based on [10]. It works by
selecting a reference point on the desired trajectory, at a predefined
distance L1 from the vehicle. Based on the angle η between the
vehicle velocity vector and a line from the vehicle to the reference
point, a lateral acceleration command is calculated as illustrated
in Figure 4. The lateral acceleration command acmd is calculated
as
acmd = 2
V2
sin η.
L1
(39)
Note that the trajectory follower will always try to move the vehi- Figure 4. Trajectory following alcle in the direction of the L1 line segment, hence pulling the vehicle gorithm
toward the desired trajectory. As the vehicle closes in on the trajectory, the vehicle is commanded to level out and approach the desired trajectory at a smaller
angle. It is also readily shown that the trajectory follower produces a lateral acceleration command
equal to the centripetal acceleration necessary to maintain a flight path on a circular segment by
13 of 18
American Institute of Aeronautics and Astronautics
Figure 5. Airplane commanded to hover at (x, y, z) = (0, 0, 1.5) m for 5 min: x–y plot of vehicle position (left),
histograms showing the radial distance error for the vehicle from (x,y) = (0,0) m (middle) and altitude error
from z = 1.5 m (right)
noting that L1 = 2R sin η. This relationship shows an inherent limit in the choices of L1 , since the
minimum radius of a circle that can be followed by the trajectory follower is Rmin = L1 /2. The
radius limitation is also made apparent by realizing that with the assumption that the vehicle is
on the desired trajectory, no reference point on a circle can be selected if the reference distance L1
is larger than the diameter of the circle that the vehicle is trying to follow.
Assuming that the above limitation on L1 is not exceeded, the choice of L1 is arbitrary. However,
due to the manner in which the trajectory follower works, a small L1 effectively makes the controller
more aggressive, whereas a larger L1 gives a smoother vehicle response. Empirical testing has shown
that the controller performance is heavily influenced by the choice of L1 . An L1 value of 2.4 m
was utilized for the trajectory following flight demonstrations presented in this paper. As with any
implementation of this tracking algorithm, it is important to have an efficient way of computing
the desired trajectory and the intersection point of the L1 line and the desired trajectory. Various
heuristics can be used to pick a target point in the case that an intersection point does not exist.
VIII.
Results
Numerous hover and level flight tests were performed using the Blade 3D and Ikarus Yak 54
Shock Flyer foam aircrafts in the RAVEN at MIT.7 First, typical results are shown in Figure 5
in which the vehicle was commanded to hold its position at (xE , yE , zE ) = (0, 0, 1.5) m for five
minutes. Figure 5 shows three plots, including a plot of the vehicle x–y location while it maintained
its position and attitude. The dashed red circle in the picture has a radius of 0.32 m. This circle
would correspond to the allowable movement in a 1.50 m wide corridor or equivalent, based on
the vehicle’s wingspan. As shown in Figure 5, the vehicle maintained its position inside this circle
during the five-minute test period. The remaining plots give histograms of the vehicle’s x-y and z
positions. The first histogram shows the radial distance error for the vehicle from (x, y) = (0, 0) m,
while the second histogram shows the vehicle’s altitude error from z = 1.5 m. These plots confirm
that the vehicle was within a 20 cm circle for over 87% of the 5 minute flight. These plots also
show that the vehicle precisely maintained its altitude (staying between 1.4 to 1.6 m) during the
entire hover test.
Next, since the airplanes are autonomous from take-off to landing, the aircraft takes off and
lands in a perching configuration. Figure 6 shows a photographed sequence of the airplane perching
on a fixed cement pole located in the middle of the RAVEN flight space in hover mode. Test flight
results for five consecutive take-off and landing sequences are shown in Figures 7 and 8. In this
test, the landing pad on the cement pole is located at (x,y,z) = (0,2.67,1.03) m in the room and
14 of 18
American Institute of Aeronautics and Astronautics
Figure 6. Autonomous landing in hover mode – perching on the fixed landing platform on a pole
Figure 7. Five consecutive landings – Airplane commanded to perch on the cement pole landing apparatus
at (x, y, z) = (0, 2.67, 1.03) m: x–y plot of vehicle position (left), z-axis position (middle), and y–z plot of vehicle
position (right)
Figure 8. Five consecutive take-offs – Airplane commanded to take-off from previous landing position on
cement pole at (x, y, z) = (0, 2.67, 1.03) m: x–y plot of vehicle position (left), z-axis position (middle), and y–z
plot of vehicle position (right)
15 of 18
American Institute of Aeronautics and Astronautics
Figure 9. Transition test – Starting from a hover, transition to and from level flight mode: x–y plot of vehicle
position (left), z-axis position (middle), and vehicle speed (right)
the vehicle was commanded to take off, hover for about 15 s and then attempt another landing. In
these plots, the vehicle begins its approach near (x,y) = (0,0,1.5) m and moves toward the landing
platform in the hover mode. As it approaches the landing location, the vehicle reduces the altitude
to be just above the landing site. Then, once the vehicle detects that it is in the proper location
with a correct heading, the vehicle descends into the landing location to await the next take-off
command. After the take-off command is given, the throttle is accelerated to pull the vehicle
directly above the landing platform before moving the vehicle away from the pole.
In Figures 7 and 8 the red diamonds at the end of each trajectory show the landing/takeoff location of the vehicle. Throughout the entire sequence, the fully-autonomous vehicle was
commanded to take off from the location where it had landed. Note that all of these take-off
locations are clustered within 0.07 m of each other. Likewise all of the landing locations are within
0.15 m of each other (in fact the first four landings are clustered within 0.07 m of each other, but the
fifth landing was on the left-hand part of the landing apparatus). Also note that the vehicle’s paths
into and off the platform are very consistent during each landing and take-off. During take-off,
there is an overshoot as the vehicle pulls itself off the platform which is a result of a safety feature
designed to ensure that the vehicle stays away from the pole.
Before attempting level flight tests with the aircraft, the vehicle’s transition logic was also
demonstrated. Figure 9 shows data from one of these transition tests in which the vehicle autonomously moved from the hover take-off location near the pole to the Northeastern area of the
flight space. The test began from the red diamond and flew south. First, note that this test occurred in under 6 s. In fact, each vehicle transition takes approximately 1.5 s to complete. As
the vehicle transitions to level flight mode the vehicle loses altitude while picking up speed. In
this test after the vehicle achieves the level flight mode (i.e., the vehicle’s calculated pitch angle
is less than 30 degrees), the vehicle is immediately commanded back to hover mode. Notice that
as the vehicle begins to pull out of the level flight state, the vehicle gains speed until the system
reduces the throttle control as the vehicle gains altitude. In addition, after the vehicle exceeds the
desired hover altitude of 1.5 m, the vehicle’s velocity reduces quickly and the vehicle’s calculated
pitch increases dramatically. In fact, the pitch angle passes 90 degrees as the vehicle tries to reduce
speed and altitude quickly to regain the hover control.
Finally, a flight demonstrating the full capabilities (hover to level flight, fly 3 laps, transition
to hover) was demonstrated. Figure 10 shows data from two of these three-lap tests. Figure 11
shows images a various stages of the flight. Once again, the vehicles autonomously moved from the
hover take-off location near the pole to the northeastern area of the flight space shown in Figure 10.
From this location, the vehicle transitioned to level flight mode, flew three laps around the pole
(as shown in the x-y location plot in Figure 10) and then transitioned back to the hover state
before attempting a landing. During the level flight in the RAVEN, the vehicle maintains a speed
16 of 18
American Institute of Aeronautics and Astronautics
Figure 10. Level flight test in constrained environment – Starting from a hover, the vehicle transitions to level
flight mode, flies three laps around the room and then transitions back to hover mode: x–y plot of vehicle
position (top), z-axis position (bottom left), and vehicle speed (bottom right). The red and blue plots represent
two separate flight tests.
Figure 11. Autonomous aircraft hover, transition to level flight and back to hover.
17 of 18
American Institute of Aeronautics and Astronautics
of approximately 6 m/s during each of the three laps. In addition, the vehicle is flying a circular
route with a radius of about 2 m. Therefore, the three laps take about 10 s from the time the
vehicle enters level flight mode to the time the vehicle transitions back to the hover state. A flight
from one of these tests can be seen at http://aerobatics.mit.edu.
IX.
Conclusion
This paper presents both vehicle models and test flight results for an autonomous fixed-wing
airplane that is designed to take-off to a hover, transition to and from traditional fixed-wing level
flight modes, and perch on a vertical landing platform in a highly space constrained environment.
The flight control for this commercially available fixed-wing model airplane are achieved using
the motion capture system available in the MIT Real-time indoor Autonomous Vehicle test ENvironment (RAVEN). These experiments represent the first known flight results in the literature
of autonomous aggressive indoor flight using a fixed-wing R/C aircraft. Future work will expand
on these test to investigate landing on moving objects in a hover and performing other aerobatic
maneuvers in the indoor test environment.
Acknowledgments
The authors would like to thank Spencer Ahrens, Brett Bethke, Luc Brunet, Eli Cohen, Daniel
Dale, Brandon Luders, and Adam Woodworth for their invaluable assistance. Research supported
by the Boeing Company, Phantom Works, Seattle and by AFOSR grant FA9550-04-1-0458.
References
1
B. L. Stevens and F. L. Lewis. Aircraft Control and Simulation, Second Edition. J. W. Wiley and Sons, 2003.
D. Biermann and R. N. Conway. Propeller Charts for the Determination of the Rotational Speed for the
Maximum Ratio of the Propeller Efficiency to the Specific Fuel Consumption. Technical report, National Advisory
Committee on Aeronautics, 1942.
3
E. N. Johnson, M. A. Turbe, A. D. Wu, S. K. Kannan, and J. C. Neidhoefer. Flight Test Results of Autonomous
Fixed-Wing UAV Transitions to and from Stationary Hover. In Proceedings of the AIAA Guidance, Navigation, and
Control Conference Exhibit, Monterey, CO, August 2006.
4
E. Frazzoli. Robust Hybrid Control for Autonomous Vehicle Motion Planning. Department of aeronautics and
astronautics, Massachusetts Institute of Technology, Cambridge, MA, June 2001.
5
A. J. Hanson. Visualizing Quaternions. Morgan Kaufman Publishers, San Francisco, CA, 2006.
6
J. B. Kuipers. Quaternions and Rotation Sequences. Princeton University Press, Princeton, NJ, 1999.
7
M. Valenti, B. Bethke, G. Fiore, J. How, and E. Feron. Indoor multi-vehicle flight testbed for fault detection,
isolation, and recovery. In Proceedings of the AIAA Guidance, Navigation, and Control Conference and Exhibit,
Keystone, CO, August 2006.
8
N. Knoebel, S. Osborne, D. Snyder, T. Mclain, R. Beard, and A. Eldredge. Preliminary modeling, control,
and trajectory design for miniature autonomous tailsitters. In Proceedings of the AIAA Guidance, Navigation, and
Control Conference and Exhibit, Keystone, CO, August 2006.
9
R. H. Stone. Control Architecture for a Tail-Sitter Unmanned Air Vehicle. In Proceedings of the 5th Asian
Control Conference, Melbourne, Australia, July 2004.
10
S. Park, J. Deyst, and J. P. How. A New Nonlinear Guidance Logic for Trajectory Tracking. In Proceedings of
the AIAA Guidance, Navigation, and Control Conference, Providence, RI, August 2004.
11
University of Sydney. T-Wing Aircraft Homepage. Avaiable at http://www.aeromech.usyd.edu.au/uav/twing/,
October 2006.
12
M. Valenti, B. Bethke, D. Dale, A. Frank, J. McGrew, S. Ahrens, J. P. How, and J. Vian. The MIT Indoor
Multi-Vehicle Flight Testbed. In Proceedings of the 2007 IEEE International Conference on Robotics and Automation
(ICRA ’07), Rome, Italy, April 2007. Video Submission.
13
W. E. Green and P. Y. Oh. A MAV That Flies Like an Airplane and Hovers Like a Helicopter. In Proceedings
of the 2005 IEEE/ASME International Conference on Advanced Intelligent Mechatronics, Monterey, CA, July 2005.
14
W. F. Chana and J. F. Coleman. World’s First VTOL Airplane Convair/Navy XFY-1 Pogo. In Proceedings
of the International Powered Lift Conference, pages 99–104, Jupiter, FL, November 1996.
2
18 of 18
American Institute of Aeronautics and Astronautics